Introduction

Farmland bird species require the edge habitats of forest/wood patches. We will investigate this main habitat by deploying autonomous recording units (ARUs) in fields near woody edges. The central area of a wood patch does not matter to us this, meaning we do not include the forest area as another dimension. We also use the proportion of permanent grassland area within 500m buffer zones and exclude the “long tail” of extreme values (>95%). From here that our main explanatory variables are:

The previous were also discretized into 7 levels to produce two additional ordinal variables.

A total of 59 sites were sampled with audiomoth-ARUs for 14 days (dawn chorus). The recordings produced were processed using the BirdNET neural network to produce a total of 2,851,110 bird species detections. # Bird detection confidence

It should be noted that BirdNET produces detections with a varying level of confidence. For the subsequent exploratory analysis, only the bird detections ranked 1 in [each of] the soft classifications were used. But even these primary detections can have a low confidence value.

The following plot shows the probability density of the confidence values of each species detections. Next to the name we also show the total amount of detections. The species are ordered from most abundant detection-wise to least.

In the original BirdNET paper the authors recommend considering primary detections with a confidence score of at least 0.5. This relatively restrictive approach led to results that match human observer performance without accounting for overlapping vocalizations during a busy dawn chorus.

So from there on we will also filter out detections with a confidence smaller than 0.5. This reduces the amount of detections from 2,851,110 to 695,651.

Sampling sites map and some first indicators

Based on the previous data we quantified the number of total number of bird detections per site and from here the estimated bird species richness. There does not seem to be any spatial clustering of any of the previous variables, not that it was expected.

The sites are presented in their true size (500m radius).

We can also map the total number of detections and richness per site on the plane defined by our two explanatory variables, grassland proportion and edge length.

Number of bird species detections

We now present some more detections per species bubble plots. A larger and darker bubble indicate more detections. The y-axis (species) is ordered from the most active species (Willow Warbler - Top) to the least active (Eurasian coot - Bottom).

The following plot shows the number of detections per species per site.

The following plot shows the number of detections per species per grassland proportion class.

The following plot shows the number of detections per species per edge length class.

Distribution of the explanatory variables per species

We created density plots from each of our two explanatory variables of interest for each species. This is presented as a (very large!) stacked ridegline plots. First for grassland proportion and then for edge length. The y-axis is ordered based on the mean of the rescpective variable. So the top most species has the largest grassland proportion/edge-length mean.

Grassland proportion per species ridgeline plot

Edge length per species ridgeline plot